Here’s what’s on our plate today:

• 🧪 Why customer data isn't becoming the AI moat investors expected.

• 📰 SpaceX's IPO blockbuster, ClickUp's layoff signal, and a new chapter in AI regulation.

• 🧠 Brain Snack: stop selling the data-moat fiction; build a real one.

• 🗳️ Poll: What is the actual AI moat in 2026?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• The flywheel is mostly fiction: The thesis borrows from the era of Google and Facebook, but privacy law, sensitive data entanglement, and standard vendor contracts collectively disable the loop before it starts.

• Legal walls before strategic ones: GDPR enforcement, including LinkedIn’s €310M fine, and the terms of OpenAI/Anthropic/Google routinely bar reusing customer data for training. The moat is contractually prohibited.

• Synthetic data kills the scarcity argument: Gartner projects that synthetic structured data will outpace real data this decade, allowing competitors to replicate statistical patterns without your users.

• Real moats look unglamorous: Harvey spent its $600M raise acquiring vLex, decades of proprietary legal content, not usage logs. Workflow integration and institutional dependency are the durable positions.

• Stakes: Enterprises committed $37B to generative AI in 2025 on a compounding-data story that may never materialize. A strategy aimed at the wrong threat is just an expensive Maginot Line.

Why customer data isn’t becoming the AI moat investors expected

There is a particular kind of military disaster that occurs when decision-makers prepare for the wrong war. France spent the 1930s building the Maginot Line: an elaborate, expensive, genuinely impressive chain of fortifications along its border with Germany. When Germany invaded in 1940, it went around it through Belgium, and the line was never breached because it never needed to be. The French weren’t wrong that fortification mattered. They were wrong about where to build it.

Something structurally similar is happening in enterprise AI right now. Across pitch decks, VC memos, and strategy meetings, one idea keeps surfacing: the ‘data moat.’ The theory is simple. The longer a company runs an AI product, the more customer interaction data it collects, the smarter the product becomes, and the harder it is for new competitors to catch up. The flywheel strengthens over time, and the advantage compounds.

That logic makes sense because it mirrors how companies like Google and Facebook built their dominance. But for many AI companies today, it may be pointing them toward the wrong competitive boundary altogether.

The theory that made sense, once

The idea of a data moat feels convincing because there are real examples where it worked. Google spent decades collecting search data that helped improve its ranking algorithms in ways competitors could not easily replicate. Facebook built an enormous social graph that became more valuable as more people joined the platform. These companies did not just have large datasets; they had data that was uniquely difficult to recreate and directly improved the product when fed back into the system.

The generative AI boom imported this logic almost unchanged. In December 2023, Greylock Partners argued that while proprietary data may provide an early advantage, the real long-term moat comes from the data customers generate as they use the product. Andreessen Horowitz initially argued in early 2023 that generative AI products lacked durable competitive advantages, but by January 2025, the firm had shifted its position, citing companies that use proprietary data to build fast-growing businesses. The broader VC consensus became clear: collect more user-interaction data, improve the model, and strengthen the moat.

However, the problem with this theory becomes apparent when you examine what customer interaction data actually consists of and what companies are legally allowed to do with it.

Where the wall leaks

In almost any enterprise vertical where AI is being deployed today, customer interactions don’t arrive as a clean, reusable training signal. They arrive wrapped in everything a company is legally obligated to protect: medical records, financial details, HR decisions, confidential contracts, and personally identifiable information. The data isn’t just sensitive in the abstract sense; it is, in the technical sense, entangled.

Companies working directly with enterprise AI are increasingly skeptical of the data moat theory. Unique.ai, which deploys AI systems in regulated financial environments, argues that most customer interaction data is tied to sensitive information that companies cannot freely reuse for training. In practice, much of what firms can safely collect is metadata: signals about how users move through software, how often they return, or where they struggle. That information may help improve the product experience, but it does not meaningfully strengthen the underlying model.

The legal environment makes the problem even harder. Under Europe’s General Data Protection Regulation, companies need a clear legal basis to use personal data for AI training. In 2024, LinkedIn was fined €310M by Ireland’s Data Protection Commission over the use of behavioral data for ad-targeting and inference systems. The case highlighted a growing tension between the idea of endlessly recycling user behavior into AI systems and the legal limits around consent and data usage.

There is also a more practical issue beneath all of this. According to Menlo Ventures, most enterprise AI tools are now bought from outside vendors rather than built internally. And the standard contracts offered by providers such as OpenAI, Anthropic, and Google usually prohibit the use of customer data to train shared models. In other words, the flywheel many investors imagine is often contractually disabled from the start.

At the same time, advances in synthetic data are weakening the scarcity argument behind the moat itself. Synthetic datasets can reproduce the statistical patterns of real-world information without containing the original data. Gartner projects that synthetic structured data will grow far faster than real structured data over the rest of the decade. If competitors can generate comparable training data artificially, then exclusive access to customer interactions becomes far less defensible as a long-term advantage.

Where real moats actually form

None of this means AI companies cannot build durable advantages. It means those advantages are emerging somewhere very different from where many investors first expected.

The clearest signal is in acquisition patterns. Harvey used much of its $600M funding to acquire vLex, a legal database business built over more than two decades. Harvey was not buying customer interaction data or usage logs. It was a large body of proprietary legal content that competitors could not easily recreate, along with deep relationships already embedded inside legal workflows. The moat came from ownership of scarce domain infrastructure, not from an AI feedback loop.

The other defensible position emerging in enterprise AI is workflow integration. The strongest AI products are increasingly the ones that become deeply woven into how an organization operates. Once an AI system is connected to internal tools, customized for company-specific processes, and trusted across teams, replacing it becomes expensive and disruptive. The advantage comes less from accumulated data and more from operational dependency, institutional familiarity, and integration depth.

The challenge is that this kind of moat is less glamorous. It does not fit neatly into the familiar flywheel story, in which products automatically become smarter through endless streams of user data. Instead, it depends on slower and more labor-intensive work: understanding industries in detail, adapting to specific workflows, and building trust over time. But in enterprise AI, that may be the only form of defensibility that consistently holds.

The lesson of the Maginot Line was never that the French failed to think strategically. The fortifications were sophisticated, expensive, and logically designed around the realities of earlier wars. The problem was that warfare evolved while the system was still being built, turning a carefully reasoned defense into protection aimed at the wrong threat.

The data moat thesis in enterprise AI risks a similar mismatch. The underlying insight was real: companies like Google and Facebook built enduring advantages by accumulating data during the internet era. But the conditions that made those moats effective do not map neatly onto modern enterprise AI, where products are often built on shared foundation models, constrained by privacy law, and governed by contracts that restrict how customer data can be reused.

Menlo Ventures estimates that enterprises will spend $37B on generative AI in 2025, much of it driven by the belief that interaction data naturally compounds into defensibility over time. Which begs the question: are enterprise AI companies still building strategy around an internet-era assumption that no longer fully matches how AI competition actually works?

Brain Snack (for Builders)

Wednesday Poll

Quick Bits, No Fluff

• SpaceX's IPO blockbuster: SpaceX's market debut drew a massive crowd of investors and easily outpaced most recent hot IPOs, cementing it as the year's most-anticipated listing.

• ClickUp's layoff signal: ClickUp's mass layoff is being read as a warning shot for SaaS more broadly, with AI-driven productivity reshaping how much headcount these companies actually need.

• A new chapter in AI regulation: A fresh BBC report details the latest moves to tighten oversight of AI systems, signaling that the regulatory honeymoon for AI companies is officially ending.

Meme Of The Day

The Toolkit

• Together AI: Cloud platform for running and fine-tuning open-source AI models at scale.

• VEED: Browser-based video editor with AI subtitles, voice cloning, and one-click translation.

• Writer: Enterprise AI writing platform with custom models trained on your brand voice.

Rate This Edition